Current browse context:
cs.LG
Change to browse by:
References & Citations
Computer Science > Machine Learning
Title: Adaptive Inertia: Disentangling the Effects of Adaptive Learning Rate and Momentum
(Submitted on 29 Jun 2020 (v1), last revised 14 Jun 2022 (this version, v11))
Abstract: Adaptive Moment Estimation (Adam), which combines Adaptive Learning Rate and Momentum, would be the most popular stochastic optimizer for accelerating the training of deep neural networks. However, it is empirically known that Adam often generalizes worse than Stochastic Gradient Descent (SGD). The purpose of this paper is to unveil the mystery of this behavior in the diffusion theoretical framework. Specifically, we disentangle the effects of Adaptive Learning Rate and Momentum of the Adam dynamics on saddle-point escaping and flat minima selection. We prove that Adaptive Learning Rate can escape saddle points efficiently, but cannot select flat minima as SGD does. In contrast, Momentum provides a drift effect to help the training process pass through saddle points, and almost does not affect flat minima selection. This partly explains why SGD (with Momentum) generalizes better, while Adam generalizes worse but converges faster. Furthermore, motivated by the analysis, we design a novel adaptive optimization framework named Adaptive Inertia, which uses parameter-wise adaptive inertia to accelerate the training and provably favors flat minima as well as SGD. Our extensive experiments demonstrate that the proposed adaptive inertia method can generalize significantly better than SGD and conventional adaptive gradient methods.
Submission history
From: Zeke Xie [view email][v1] Mon, 29 Jun 2020 05:21:02 GMT (789kb,D)
[v2] Tue, 30 Jun 2020 00:51:17 GMT (789kb,D)
[v3] Fri, 17 Jul 2020 09:27:31 GMT (789kb,D)
[v4] Sun, 16 Aug 2020 07:11:34 GMT (791kb,D)
[v5] Mon, 24 Aug 2020 04:59:13 GMT (812kb,D)
[v6] Thu, 27 Aug 2020 04:01:46 GMT (892kb,D)
[v7] Wed, 28 Oct 2020 08:42:20 GMT (1073kb,D)
[v8] Tue, 24 Nov 2020 05:03:01 GMT (1074kb,D)
[v9] Sun, 7 Feb 2021 11:53:48 GMT (1797kb,D)
[v10] Mon, 23 May 2022 02:09:30 GMT (1855kb,D)
[v11] Tue, 14 Jun 2022 15:25:21 GMT (1872kb,D)
Link back to: arXiv, form interface, contact.